JMIR Publications

ABSTRACT

Background:

Large language models can sort pathology reports to identify breast cancer molecular subtypes. Such LLM models can distinguish breast cancer subtypes like Luminal A, Luminal B, HER2-enriched, and triple-negative breast cancer based on the presence or absence of key receptors like estrogen, progesterone, and human epidermal growth factor receptors, enabling more precise and targeted therapeutic decisions. LLM models also assist in accurate data extraction of oncology information, which is useful in identifying subtypes of breast cancer.

Objective:

This paper proposes a new framework based on LLM models to automate breast cancer subtype triage from pathology reports in an outpatient department setting, utilizing oncology features from breast cancer pathology reports to drive subtype triage in breast cancer patients. Key receptor status, tumor characteristics, and other biological profiles, such as tumor size, tumor grade, and tumor histology, are included in this LLM model to capture the biological profile of breast cancer patients.

Methods:

A total of 532 breast cancer pathology reports from TCGA are included in this study to evaluate the LLM model’s performance in triaging molecular subtypes of breast cancer patients. This research paper utilizes a variety of lightweight generative models, such as GPT-4.0 Mini, GPT-4.1 Nano, GPT-5 Nano, and Gemini 2.5 Flash Lite, to process breast cancer pathology reports and test their performance in a chain-of-thought, zero-shot, and few-shot setting.

Results:

The LLM model extracted the oncology features successfully with the highest accuracy achieved by GPT-4oMini at 89.21% for chain-of-thought prompting. Higher accuracy was achieved in the extraction of biomarkers, with tumor size having lower accuracy. The GPT-5 Nanomodel was able to identify the cases of TNBC in 44 out of the 53 cases at ≈ 83%.

Conclusions:

The LLMmodels, with the aid of structured prompting, can be used for the precise extraction of oncology features. The GPT-5 Nano model was able to show high accuracy in the extraction of the features of TNBC.